Post on 13-Jul-2020
Econ 219B
Psychology and Economics: Applications
(Lecture 10)
Stefano DellaVigna
April 1, 2015
Outline
1. Projection Bias
2. Non-Standard Decision-Making
3. Attention: Introduction
4. Attention: Simple Model
5. Attention: eBay Auctions
6. Attention: Taxes
7. Attention: Left Digits
8. Attention: Financial Markets
9. Methodology: Portfolio Methodology
1 Projection Bias
• Beliefs systematically biased toward current state
• Read-van Leeuwen (1998):— Office workers choose a healthy snack or an unhealthy snack
— Snack will be delivered a week later (in the late afternoon).
— Two groups: Workers are asked
∗ when plausibly hungry (in the late afternoon) — 78 percent chosean unhealthy snack
∗ when plausibly satiated (after lunch).— 42 percent choose unhealthysnack
• Gilbert, Pinel, Wilson, Blumberg, and Wheatly (1999):— individuals under-appreciate adaptation to future circumstances —Projection bias about future reference point
— Subjects forecast happiness for an event
— Compare predictions to responses after the event has occurred
— Thirty-three current assistant professors at the University of Texas(1998) forecast that getting tenure would significantly improve theirhappiness (5.9 versus 3.4 on a 1-7 scale).
— Difference in rated happiness between 47 assistant professors that wereawarded tenure by the same university and 20 that were denied tenureis smaller and not significant (5.2 versus 4.7).
— Similar results as function of election of a Democratic of Republicanpresident, compared to the realized ex-post differences.
• Projection bias. (Loewenstein, O’Donoghue, and Rabin (2003)— Individual is currently in state 0 with utility
¡ 0
¢— Predict future utility in state
— Simple projection bias:
( ) = (1− ) ( ) + ³ 0
´— Parameter is extent of projection bias — = 0 implies rationalforecast
• Notice: People misforecast utility , not state ; however, same results ifthe latter applies
• Conlin-O’Donoghue-Vogelsang (2006)• Purchasing behavior: Cold-weather items• Main Prediction:— Very cold weather
— — Forecast high utility for cold-weather clothes
— — Purchase ‘too much’
— — Higher return probability
• Denote temperature at Order time as and temperature at Return timeas
• Predictions:1. If = 0 (no proj. bias), P[R|O] is independent of and 2. If 0 (proj. bias), P[R|O] 0 and P[R|O] 0
• Purchase data from US Company selling outdoor apparel and gear— January 1995-December 1999, 12m items— Date of order and date of shipping + Was item returned— Shipping address
• Weather data from National Climatic Data Center— By 5-digit ZIP code, use of closest weather station
• Items:— Parkas/Coats/Jackets Rated Below 0F— Winter Boots— Drop mail orders, if billing and shipping address differ, 9 items or-dered, multiple units same item, low price
— No. obs. 2,200,073
• Note: Probability of return fairly high, Delay between order and receipt4-5 days
• Main estimation: Probit (|) = Φ (+ + +)
• Main finding: 0
— Warmer weather on order date lowers probability of return
— Magnitude:
— This goes against standard story: If weather is warmer, less likely youwill use it — Return it more
— Projection Bias: Very cold weather — Mispredict future utility —Return the item
• Second finding: ≈ 0— Warmer weather on (predicted) return does not affect return
— This may be due to the fact that do not observe when return decisionis made
• Similar estimates for linear probability model with household fixed effects
• Simple structural model: Estimates of projection bias around .3-.4
• Busse, Pope, Pope, Silva-Risso (2013): Evidence from car purchasesand house purchases
• Projection bias:— Convertible looks particularly attractive on a hot day
— 4-wheel drive attractive on snowy day
— House with pool higher selling price on hot day
• Strong evidence in the data
2 Non-Standard Decision-Making
• First part of class: Non-standard preferences (|):— Over time (present-bias)
— Over risk (reference-dependence)
— Over social interactions (social preferences)
• And Non-Standard Beliefs ()— About skill (overconfidence)
— Updating (law of small numbers)
— About preferences (projection bias)
• Third category: Standard (|) and () — Still, non-standard deci-sions
• Sub-categories— Limited attention
— Framing
— Menu effects
— Persuasion and social pressure
— Emotions
— Happiness
— Mental Accounting
• This in turn often leads to non-standard beliefs e ()
3 Attention: Introduction
• Attention as limited resource• Psychology Experiments: Dichotic listening (Broadbent, 1958)— Hear two messages:
∗ in left ear∗ in right ear
— Instructed to attend to message in one ear
— Asked about message in other ear — Cannot remember it
— More important: Asked to rehearse a number (or note) in their head— Remember much less the message
• Attention clearly finite
• How to optimize given limited resources?— Satisficing choice (Simon, 1955 — Conlisk, JEL 1996)
— Heuristics for solving complex problems (Gabaix-Laibson, 2002; Gabaixet al., 2003)
• In a world with a plethora of stimuli, which ones do agents attend to?• Psychology: Salient stimuli (Fiske-Taylor, 1991) — Not very helpful
• Probably, no general rule — Inattention along many dimensions
• Does this apply to high-stakes items?• Event of economic importance: Huberman-Regev (JF, 2001)• Timeline:— October-November 1997: Company EntreMed has very positive earlyresults on a cure for cancer
— November 28, 1997: Nature “prominently features;” New York Timesreports on page A28
— May 3, 1998: New York Times features essentially same article as onNovember 28, 1997 on front page
— November 12, 1998: Wall Street Journal front page about failed repli-cation
• In a world with unlimited arbitrage...
• In reality...
• At least two interpretations:1. Limited attention initially + Catch up later
2. Full incorporation initially + Overreaction later
• Persistence for 6 months suggests (1) more plausible• Other interpretations:— Focal point
— non-Bayesian inference
4 Attention: Simple Model
• Simple model (DellaVigna JEL 2009)• Consider good with value (inclusive of price), sum of two components: = +
1. Visible component
2. Opaque component
• Inattention— Consumer perceives the value = + (1− )
— Degree of inattention , with = 0 standard case
— Interpretation: each individual sees , but processes it only partially, tothe degree
• Alternative model:— share on individuals are inattentive, 1− attentive —
— Models differ where not just mean, but also max/min matter (Ex.:auctions)
• Inattention is function of:— Salience ∈ [0 1] of with 0 0 and (1 ) = 0
— Number of competing stimuli : = () with 0 0 (Broad-bent)
• Consumer demand [ ] with 0[] 0 for all
• Model suggests three strategies to identify the inattention parameter :
1. Compute response of to change in — compare = (1− )
to = 1 (Hossain-Morgan (2006), Chetty-Looney-Kroft (2009),Lacetera-Pope-Sydnor (2012), Cohen-Frazzini (2011))
2. Examine the response of to an increase in the salience , =−0: differs from zero? (Chetty et al. (2009))
3. Vary competing stimuli , = −0 : differs from zero?(DellaVigna-Pollet (2009) and Hirshleifer-Lim-Teoh (2009))
• Key element: identify opaque information
• Two caveats:
— Measuring salience of information is subjective – psychology experi-ments do not provide a general criterion
— Inattention can be rational or not.
∗ Can rephrase as rational model with information costs
∗ However, opaque information is publicly available at a zero or smallcost (for example, earnings announcements news)
∗ Rational interpretation less plausible
5 Attention: eBay Auctions
• Hossain-Morgan (2006). Inattention to shipping cost
• Setting:— is value of the object
— negative of the shipping cost: = −— Inattentive bidders bid value net of the (perceived) shipping cost: ∗ = − (1− ) (2nd price auction)
— Revenue raised by the seller: = ∗ + = +
— Hence, $1 increase in the shipping cost increases revenue by dollars
— Full attention ( = 0): increases in shipping cost have no effect onrevenue
• Field experiment selling CD and XBoxs on eBay— Treatment ‘LowSC’ [A]: reserve price = $4 and shipping cost = $0
— Treatment ‘HighSC’ [B]: reserve price = $01 and shipping cost = $399
— Same total reserve price = + = $4
— Measure effect on total revenue probability of sale
• Predictions:— Standard model: = 0 = — =
— Inattention: = —
• Strong effect: − = $261 —Inattention = 2614 = 65
• Smaller effect for XBox: − = $071 — Inattention = 0714 =18
• Pooling data across treatments: in 16 out of 20 cases —Significant difference
• Similar treatment with high reserve price:— Treatment ‘LowSC’ [C]: reserve price = $6 and shipping cost = $2
— Treatment ‘HighSC’ [D]: reserve price = $2 and shipping cost = $6
• No significant effect for CDs (perhaps reserve price too high?): − =
−29 — Inattention = −294 = −07
• Large, significant effect for XBoxs: − = 411 — Inattention = 4114 = 105
• Overall, strong evidence of partial disregard of shipping cost: ≈ 5
• Inattention or rational search costs
6 Attention: Taxes
• Chetty et al. (AER, 2009): Taxes not featured in price likely to beignored
• Use data on the demand for items in a grocery store.
• Demand is a function of:
— visible part of the value , including the price
— less visible part (state tax −)— = [ − (1− ) ]
• Variation: Make tax fully salient ( = 1)
• Linearization: change in log-demand∆ log = log [ − ]− log [ − (1− ) ] =
= − ∗0 [ − (1− ) ] [ − (1− ) ]
= − ∗
— is the price elasticity of demand
— ∆ log = 0 for fully attentive consumers ( = 0)
— This implies = −∆ log( ∗ )
• Part I: field experiment— Three-week period: price tags of certain items make salient after-taxprice (in addition to pre-tax price).
• Compare sales to:
— previous-week sales for the same item
— sales for items for which tax was not made salient
— sales in control stores
— Hence, D-D-D design (pre-post, by-item, by-store)
• Result: average quantity sold decreases (significantly) by 2.20 units relativeto a baseline level of 25, an 8.8 percent decline
• Compute inattention:— Estimates of price elasticity : −159— Tax is 07375
— = −(−088)(−159 ∗ 07375) ≈ 75
• Additional check of randomization:— Generate placebo changes over time in sales
— Compare to observed differences
— Use Log Revenue and Log Quantity
• Non-parametric p-value of about 5 percent
• Part II: Panel Variation— Compare more and less salient tax on beer consumption
— Excise tax included in the price
— Sales tax is added at the register
— Panel identification: across States and over time
— Indeed, elasticity to excise taxes substantially larger — estimate of theinattention parameter of = 94
• Substantial consumer inattention to non-transparent taxes
7 Attention: Left Digits
• Are consumers paying attention to full numbers, or only to more salientdigits?
• Classical example: =$5.99 vs. =$6.00
• Consumer inattentive to digits other than first, perceive = 5 + (1− ) 99
= 6
− = 01 + 99
• Optimal Pricing at 99 cents
• Indeed, evidence of 99 cents effect in pricing at stores
• Shlain and Brot-Goldberg (2014):
• Write down predicted pricing with left-digit inattention— Not only bunching at 99 cents
— Also no pricing at 0, 10 cents
— Pricing at 49 cents, 59 cents, etc.
• Examine a change in Israel which eliminates the second digit— Most prices switch to 90 cents as model predicts
— Some prices swtich to 0 cents — a puzzle!
— Over time, the 0 cents disapper... a victory for the model
• Ashton (2014): Re-analysis of Chetty et al. data— Show that effect on sales is concentrated to cases in which first digtchanges
∗ Not much effect if adding tax raises price from 3.50 to 3.80∗ Effect is adding tax raises price from 3.99 to 4.30
— Compute DDD for Shifting digit and Rigid digit
— Effect is entirely due to Shifting Digit
• Lacetera, Pope, and Sydnor (AER 2012). Inattention in Car Sales
• Sales of used cars —Odometer is important measure of value of car
• Suppose perceived value of car is
= −
• Perceived mileage is = ( 10) + (1− )mod( 10)
• Model predicts jump in value at 10k discontinuity of
−10while slope is
− (1− )
• Can estimate inattention parameter : Jump/Slope gives (1− )
• Data set— 27 million wholesale used car auctions
— January 2002 to September 2008
— Buyer: Used car dealer
— Seller: car dealer or fleet/lease
— Continuous mileage displayed prominently on auction floor
• Result: Amazing resemblance of data to theory-predicted patterns: jumpat 10k mark
— Sizeable magnitudes: $200
• If discontinuity, expect smaller jumps also at 1k mileage points
• Structural estimation of limited attention parameter can be done withDelta method or with NLS
— Structural estimation can be from OLS
— Estimate = 033 (001) for dealers, = 022 (001) for lease
— Remarkable precision in the estimate of inattention
— Consistent with other evidence, but much more precise
• Who does this inattention refer to?1. Auction buyers are biased — But these are used car re-sellers
2. Ultimate car buyers are biased — Auction buyers incorporate it in bids
• Provide some evidence on experience of used car buyers:1. Hyp. 1 implies more experienced buyers will not buy at 19,990
2. Hyp. 2 implies more experienced buyers will indeed buy at 19,990
• Behavioral IO:— Biases of consumers
— Rational firms respond to it, altering transaction price
• Would like more direct evidence: Do ultimate car buyers display bias?• Busse, Lacetera, Pope, Silva-Risso, Sydnor (AER P&P 2013)— Data from 16m transaction of used cars
— Information on sale price
— Same time period
— Is there similar pattern? Yes
• Similar estimate of inattention for auction buyers and ultimate buyers
• Heterogeneity by income (at ZIP level)? Some
8 Attention: Financial Markets I
• Is inattention limited to consumers?
• Finance: examine response of asset prices to release of quarterly earningsnews
• Setting:— Announcement a time
— is known information about cash-flows of the company
— is new information in earnings announcement
— Day − 1: company price is −1 =
— Day :
∗ company value is +
∗ Inattentive investors: asset price responds only partially to thenew information: = + (1− ) .
— Day + 60: Over time,price incorporates full value: +60 = +
• Implication about returns:— Short-run stock return equals = (1− )
— Long-run stock return , instead, equals =
— Measure of investor attention: ()() = (1− ) —Test: Is this smaller than 1?
— (Similar results after allowing for uncertainty and arbitrage, as long aslimits to arbitrage – see final lectures)
• Indeed: Post-earnings announcement drift (Bernard-Thomas, 1989): Stockprice keeps moving after initial signal
• Inattention leads to delayed absorption of information.
• DellaVigna-Pollet (JF 2009)— Estimate ()() using the response of returns tothe earnings surprise
— : returns in 2 days surrounding an announcement
— : returns over 75 trading days from an announcement
• Measure earnings news :
= −
−1— Difference between earnings announcement and consensus earningsforecast by analysts in 30 previous days
— Divide by (lagged) price −1 to renormalize
• Next step: estimate
• Problem: Response of stock returns to information is highly non-linear
• How to evaluate derivative?
9 Methodology: Portfolio Methodology
• Economists’ approach:— Make assumptions about functional form — Arctan for example
— Do non-parametric estimate — kernel regressions
• Finance: Use of quantiles and portfolios (explained in the context ofDellaVigna-Pollet (JF 2009))
• First methodology: Quantiles— Sort data using underlying variable (in this case earnings surprise )
— Divide data into equal-spaced quantiles: = 10 (deciles), = 5
(quintiles), etc
— Evaluate difference in returns between top quantiles and bottom quan-tiles: −1
• This paper:— Quantiles 7-11. Divide all positive surprises
— Quantiles 6. Zero surprise (15-20 percent of sample)
— Quantiles 1-5. Divide all negative surprise
• Notice: Use of quantiles "linearizes" the function
• Delayed response − (post-earnings announcement drift)
• Inattention:— To compute use 11−1 = 00659 (on non-Fridays)
— To compute use 11−1 = 01210 (on non-Fridays)— Implied investor inattention: ()() = (1− ) =544 — Inattention = 456
• Is inattention larger when more distraction?
• Weekend as proxy of investor distraction.— Announcements made on Friday: ()() is 41 per-cent — ≈ 59
• Second methodology: Portfolios— Instead of using individual data, pool all data for a given time period into a ‘portfolio’
— Compute average return for portfolio over time
— Control for Fama-French ‘factors’:
∗ Market return ∗ Size ∗ Book-to-Market
∗ Momentum
∗ (Download all of these from Kenneth French’s website)— Regression:
= + +
— Test: Is significantly different from zero?
• Example in DellaVigna-Pollet (2009)
— Each month portfolio formed as follows: (11 − 1 )− (11− −1− )
— Returns (3-75) -Differential drift between Fridays and non-Fridays
• Intercept = 0384 : monthly returns of 3.84 percent from this strategy
10 Attention: Financial Markets II
• Cohen-Frazzini (JF 2011) — Inattention to subtle links
• Suppose that you are a investor following company A
• Are you missing more subtle news about Company A?
• Example: Huberman and Regev (2001) — Missing the Science article
• Cohen-Frazzini (2011) — Missing the news about your main customer:— Coastcoast Co. is leading manufacturer of golf club heads
— Callaway Golf Co. is leading retail company for golf equipment
— What happens after shock to Callaway Co.?
• Data:— Customer- Supplier network — Compustat Segment files (RegulationSFAS 131)
— 11,484 supplier-customer relationships over 1980-2004
• Preliminary test:— Are returns correlated between suppliers and customers?
— Correlation 0.122 at monthly level
• Computation of long-short returns— Sort into 5 quintiles by returns in month of principal customers,
— By quintile, compute average return in month + 1 for portfolio ofsuppliers +1:
1+1
2+1
3+1
4+1
5+1
— By quintile , run regression
+1 = + +1 + +1
— +1 are the so-called factors: market return, size, book-to-market,and momentum (Fama-French Factors)
— Estimate gives the monthly average performance of a portfolio inquintile
— Long-Short portfolio: 5 − 1
• Results in Table III: Monthly abnormal returns of 1.2-1.5 percent (huge)
• Information contained in the customer returns not fully incorporated intosupplier returns
• Returns of this strategy are remarkably stable over time
• Can run similar regression to test how quickly the information is incorpo-rated
— Sort into 5 quintiles by returns in month of principal customers,
— Compute cumulative return up to month k ahead, that is, −+— By quintile , run regression of returns of Supplier:
−+ = + + + +1
— For comparison, run regression of returns of Customer:
−+ = + + + +1
• For further test of inattention, examine cases where inattention is morelikely
• Measure what share of mutual funds own both companies: COMOWN
• Median Split into High and Low COMOWN (Table IX)
11 Next Lecture
• Framing
• Menu Effects:— Choice Avoidance
— Preference for Familiar
— Preference for Salient
— Confusion
• Persuasion
• Emotions: Mood